1.1K ·GitHub
What is this?
It is 'Cold Brew' (CHA), a jailbreak prompt workbench that packages prompt-injection techniques to break the safety guardrails ('破甲越狱') of six chat models — GPT-6 Astra/Codex, Claude Code, Grok 4.6, DeepSeek v4 Pro, GLM 5.3 and Gemini — for security research.
It is 'Cold Brew' (CHA), a jailbreak prompt workbench that packages prompt-injection techniques to break the safety guardrails ('破甲越狱') of six chat models — GPT-6 Astra/Codex, Claude Code, Grok 4.6, DeepSeek v4 Pro, GLM 5.3 and Gemini — for security research.Is this a set of jailbreak prompt files, or a full desktop app?Both: each model gets its own jailbreak prompt pack plus a desktop seat with preview, run, verify and restore actions that writes the pack into that model's local instruction layer.
Both: each model gets its own jailbreak prompt pack plus a desktop seat with preview, run, verify and restore actions that writes the pack into that model's local instruction layer.How do I get it running?Build the Windows portable package with npm run pack:win, then use each model's desktop seat to preview, run, verify or restore its prompt pack; seat injection details are documented in docs/SEAT-PACKS.md.
Build the Windows portable package with npm run pack:win, then use each model's desktop seat to preview, run, verify or restore its prompt pack; seat injection details are documented in docs/SEAT-PACKS.md.Do I need Node.js before building the packs?Yes — the only documented build command is an npm script, so Node.js and npm are required to produce the Windows portable package.
Yes — the only documented build command is an npm script, so Node.js and npm are required to produce the Windows portable package.Is there a safety net if a jailbreak pack breaks a model session?Yes — the original prompt is backed up under each seat's cha-backups directory before injection, 'verify' checks the marked instruction blocks, and 'restore' rolls everything back.
Yes — the original prompt is backed up under each seat's cha-backups directory before injection, 'verify' checks the marked instruction blocks, and 'restore' rolls everything back.What does the Cold Brew visual branding look like?A dark manga-style identity built from black, bone-white and crimson tones with broken-panel comic layouts, replacing the project's older blue-green tech-panel look across its page, desktop app and community channels.
A dark manga-style identity built from black, bone-white and crimson tones with broken-panel comic layouts, replacing the project's older blue-green tech-panel look across its page, desktop app and community channels.Which six models does it provide jailbreak seats for?GPT-6 Astra/Codex, Claude Code, Grok 4.6, DeepSeek v4 Pro, GLM 5.3 and Gemini, each kept on what the README calls 'the same black-and-white negative' seat layout.
GPT-6 Astra/Codex, Claude Code, Grok 4.6, DeepSeek v4 Pro, GLM 5.3 and Gemini, each kept on what the README calls 'the same black-and-white negative' seat layout.Is this an officially sanctioned tool or a research exploit kit?It is explicitly framed as a 破甲越狱 (armor-breaking / jailbreak) security-research toolkit rather than a sanctioned product, and it ships with no clear license (GitHub reports NOASSERTION), so usage risk and legality fall on the operator.
It is explicitly framed as a 破甲越狱 (armor-breaking / jailbreak) security-research toolkit rather than a sanctioned product, and it ships with no clear license (GitHub reports NOASSERTION), so usage risk and legality fall on the operator.What are similar projects?Ai-jailbreak, dsh-infinite-gen-5, L1B3RT4S, Spiritual-Spell-Red-Teaming, AI-Infra-Guard, awesome-agent-skills-security, dsh-purge, 0
Ai-jailbreak, dsh-infinite-gen-5, L1B3RT4S, Spiritual-Spell-Red-Teaming, AI-Infra-Guard, awesome-agent-skills-security, dsh-purge, 0Which topics is it in?llm-jailbreak, prompt-injection, red-teaming, llm-security, glm, jailbreak, security-research, ai-security, grok, prompt-engineering, cybersecurity, deepseek
llm-jailbreak, prompt-injection, red-teaming, llm-security, glm, jailbreak, security-research, ai-security, grok, prompt-engineering, cybersecurity, deepseek